Generative Occupancy Fields for 3D Surface-Aware Image Synthesis

Authors: Xudong XU, Xingang Pan, Dahua Lin, Bo Dai

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments. Comparison with baselines. To validate the effectiveness of GOF, we compare it with two representative GRAF methods, namely GRAF [7] and pi-GAN. Firstly, Fig. 3 demonstrates the qualitative comparison between these three methods... Table 1: Quantitative results (128 128 px) on BFM, Celeb A and Cats datasets, on three metrics Fréchet Inception Distance (FID), Inception Score (IS) and the weighted variance of sampled depth Σti( 10 4).
Researcher Affiliation Academia Xudong Xu Xingang Pan Dahua Lin Bo Dai CUHK Sense Time Joint Lab, The Chinese University of Hong Kong Max Planck Institute for Informatics S Lab, Nanyang Technological University {xx018, dhlin}@ie.cuhk.edu.hk xpan@mpi-inf.mpg.de bo.dai@ntu.edu.sg
Pseudocode No The paper describes its method through text and mathematical equations, but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/Sheldon Tsui/GOF_Neur IPS2021.
Open Datasets Yes To assess our method comprehensively, we conduct experiments on three datasets, namely Celeb A [44], BFM [45], and Cats [46].
Dataset Splits No The paper states that it learns from unposed images and mentions using a 'test split of BFM' for evaluating a separate CNN trained on generated data, but it does not specify explicit train/validation/test splits for the GOF model's training process.
Hardware Specification Yes To compare the efficiency straightforwardly, we estimate the rendering speed of 256 256 images for both pi-GAN and GOF on a single Intel Xeon(R) CPU.
Software Dependencies No The paper provides implementation details for its model parameters but does not list any specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes Unless stated otherwise, in all experiments we set N, the number of points sampled for rendering, to 12, and set M, the number of bins used in root-finding, to 12... In practice, the number of iterations is set to ms = 3 times... In practice, we empirically set τ as 0.5.